fix: a_until_dialog_end breaks after a single A press (wrong dict key)

[golldyck]

The a_until_dialog_end action is supposed to press A every 30 frames until the dialog clears (max 300), but it exits after a single press.

Cause

In _execute_action (pokemon_agent/server.py), the loop checks:

state = _get_state_dict()
if not state.get("dialog_active", False):
    break

Dialog state is nested. _get_state_dict() returns state["dialog"] = reader.read_dialog(), which is {"active": bool, ...}. There is no top-level "dialog_active" key. So .get("dialog_active", False) always returns False, not False is True, and the loop breaks on the first iteration.

The rest of the codebase already reads this flag from the nested location:

• autopilot.py:95 uses (state.get("dialog") or {}).get("active")
• dashboard/history.py:240 uses state.get("dialog", {}).get("active")

Effect

a_until_dialog_end presses A exactly once (30 frames) instead of looping. Multi-box NPC dialogs and long text are never advanced by this action, which defeats the purpose of the command.

Fix

Read the nested flag with the same defensive pattern used elsewhere. This also handles dialog being None, e.g. FireRed where read_dialog is not yet implemented:

if not (state.get("dialog") or {}).get("active", False):
    break

Tests

Adds test_dialog_action.py with a fake emulator/reader (no ROM needed):

• loops until the dialog clears: 4 A presses across 3 active checks plus the clearing one
• respects the 300-frame cap: stops at 10 presses when the dialog never clears
• stops immediately when no dialog: 1 press

The first two fail on main (1 press each, proving the single-press bug) and pass with this change.